Deepak Pathak, a new faculty member at Carnegie Mellon University’s Robotics Institute, discusses ways to help learning systems generalize beyond training, in a Robotics Institute seminar.
Pathak argues that machine learning systems often rely on training and test data coming from similar distributions. He illustrates the gap with image classifiers that perform well on standard datasets but mislabel objects in ordinary YouTube video, and with robots whose learned skills do not transfer easily to new tasks.
Pathak contrasts imitation learning, which needs demonstrations for each task, with reinforcement learning, which can be impractical when rewards are rare. Inspired by childhood exploration, his team uses prediction error or disagreement among models to guide goal-free exploration. On a robot with a few objects, he reports meaningful behavior after five to six hours and a few thousand samples; the collected data can support later tasks such as knot tying and navigation.
Pathak says many robot learning setups use slow, step-by-step actions that struggle with dynamic tasks such as catching a ball. His team’s neural dynamic policies have a network output the parameters of a dynamical system, producing smooth trajectories while learning from rewards or demonstrations. He reports gains in simulation and says real-world transfer was in progress.
Pathak describes policies built from local controllers for each motor, which communicate through learned messages. A shared policy trained across 50 simulated robots, with at least 10 motors each, transferred to new robot shapes, though performance varied and some configurations failed. He also discusses early work on automatic simulator tuning from images and 3D scene representations that wrap a learned sheet around a scene to generate new views from one image.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Deepak Pathak and Abhinav Gupta, Skild co-founders and robotics researchers, explain Skild’s general-purpose robot brain on the NVIDIA AI Podcast with host Noah Kravitz.
At GTC, NVIDIA Omniverse host and panelists Pras Velicodi (Agility Robotics), Ashok Elluswamy (Tesla), Chelsea (Physical Intelligence and Stanford), Arnaud Rubaire (Exon Robotics), Deepak Pathak (Scaled AI and Carnegie Mellon), and Russ discuss data, simulation, robot brains, and deploying humanoids at scale.
